Observable-Space Reconstruction

• Observable-space reconstruction is a framework that leverages directly measured data—incorporating detection, masking, and visibility—to constrain inverse problems.
• It avoids unsupported extrapolation by formulating reconstructions within the observable domain, ensuring that only detectable data informs the results.
• Applications span gravitational-wave analysis, CMB lensing, redshift-space mapping, and 3D vision while addressing calibration, stability, and incomplete data challenges.

Observable-space reconstruction denotes a class of inverse methods that formulate reconstruction directly in the domain of what is measured, detected, or visibly supported by data: detector-observable source parameters in gravitational-wave astronomy, local temperature correlations on the sky for CMB lensing, redshift-space density fields in large-scale structure, feasible visual-hull volumes in single-image 3D reconstruction, and observable trace sequences in Floquet tomography (Toubiana et al., 17 Jul 2025, Bucher et al., 2010, Zhu et al., 2017, Ryu et al., 2023, Kamata, 23 May 2026). Taken together, these uses suggest a methodological stance rather than a single canonical algorithm: the reconstruction target is constrained from the outset by detectability, masking, visibility, or observable algebra, instead of being inferred only after deconvolving to an intrinsic latent space.

1. Scope, terminology, and recurring structure

Across the cited literature, the “observable space” is the measurement domain in which the forward model is well defined. In gravitational-wave population inference, it is the distribution $$p_o(\theta\mid\Lambda)=p(\theta\mid\mathrm{det},\Lambda)$$ obtained by conditioning an astrophysical population on detection. In CMB lensing, it is a set of local quadratic combinations of nearby temperature pixels. In redshift-space reconstruction, it is the observed coordinate $s$ rather than the real-space Eulerian position $x$. In single-image 3D reconstruction, it is the voxel set consistent with predicted depth and silhouette. In Floquet tomography, it is the sequence $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$ generated by repeated measurements of a fixed observable (Toubiana et al., 17 Jul 2025, Bucher et al., 2010, Zhu et al., 2017, Ryu et al., 2023, Kamata, 23 May 2026).

Domain	Observable space	Reconstruction target
Gravitational waves (Toubiana et al., 17 Jul 2025)	Detector-observable parameters $\theta$	Observable population $p_o(\theta\mid\Lambda)$
CMB and LSS (Bucher et al., 2010, Zhu et al., 2017, Chung et al., 2 Jul 2025)	Real-space pixel pairs, redshift space, ISW line integrals	Lensing fields, displacement potential, gravitational perturbations
Vision and geometry (Li et al., 2021, Ryu et al., 2023, Lee et al., 2024)	Partial surfaces, carved feasible voxels, partial object scans	Missing geometry, motion fields, completed objects
Dynamical and operator systems (Zhang et al., 2021, Kamata, 23 May 2026)	Motion traces $p(t)$, observable trace sequences $\zeta_n^{(O)}$	Latent phase-space descriptors, spectral data, visible representatives

A recurring structure is apparent. First, a forward model maps latent structure to measured data. Second, detectability or visibility is built into the inference objective rather than treated as an afterthought. Third, reconstruction is limited to a visible region or visible subspace, often with an explicit statement about what remains unidentifiable. This suggests that observable-space reconstruction is especially attractive when deconvolution to an intrinsic space would require extrapolation outside the support of the instrument or model.

2. Direct statistical inference in detector-observable parameter space

In gravitational-wave astronomy, observable-space reconstruction is formulated by starting from an astrophysical population $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$, defining the detection probability

$$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$

and then conditioning on detection to obtain

$s$0

With $s$1, the observable differential rate is $s$2 (Toubiana et al., 17 Jul 2025).

The key likelihood rewrites hierarchical inference directly in terms of the observable population: $s$3 Operationally, posterior samples $s$4 from single-event parameter estimation are reweighted by $s$5 and by the parameter-estimation prior $s$6, yielding a Monte Carlo approximation of each event integral. A typical implementation precomputes $s$7 with an emulator or injection campaign, chooses a parametric or nonparametric model for $s$8, and samples the hyperparameters $s$9 with nested sampling or MCMC (Toubiana et al., 17 Jul 2025).

The central conceptual point is that this framework does not eliminate selection effects; it incorporates them differently. The denominator $x$0 remains essential, and the framework assumes that $x$1 is known and numerically accurate above a chosen floor such as $x$2. In the application to 59 O3 BBH events, using LVK public posteriors with the “Power Law + Peak” priors and a detection-probability emulator trained on O3 injections in realistic noise, the inferred observable redshift rate $x$3 tracks the LVK reweighted result and lies within $x$4 credible intervals of a fiducial population-synthesis model; the observable $x$5 distribution also agrees with both the LVK reweighted inference and the population-synthesis prediction once selection effects are applied (Toubiana et al., 17 Jul 2025). A common misconception is therefore that “observable-space” means “selection-free”; the formalism explicitly rejects that interpretation.

3. Cosmological reconstructions in measured coordinates

In CMB lensing, real-space reconstruction replaces the customary nonlocal harmonic-space quadratic estimator by compact local estimators built from nearby pixel pairs. The reconstructed fields are the dilatation $x$6 and two shear components $x$7 and $x$8, directly related to the Hessian of the lensing potential through $x$9. The minimum-variance dilatation estimator takes the form

$\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$0

with analogous shear estimators defined by kernels $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$1 and $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$2 (Bucher et al., 2010). The practical motivation is the cut and masked sky: real-space estimators are faster to implement and naturally accommodate a galactic cut and small excisions. Quantitatively, at $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$3 the combined dilatation plus longitudinal-shear estimator has nearly the same noise as the full quadratic estimator, while at $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$4 the loss is only $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$5 in $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$6 for Planck-like noise (Bucher et al., 2010).

Nonlinear reconstruction of redshift space pushes the same logic into large-scale structure. The observed position is

$\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$7

where $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$8, and the reconstruction introduces potential isobaric coordinates $\zeta_n^{(O)}=\mathrm{Tr}(OM^n)$9 satisfying

$\theta$0

The reconstructed density is then $\theta$1 (Zhu et al., 2017). Numerically, a moving-mesh multigrid solver enforces constant mass per cell. The reconstructed anisotropic field correlates with the linear initial conditions far beyond the unreconstructed redshift-space density: the monopole cross-correlation satisfies $\theta$2 out to $\theta$3 and $\theta$4 out to $\theta$5, whereas the unreconstructed field reaches those thresholds only to $\theta$6 and $\theta$7, respectively. Expressed as “number of linear modes,” the gain is $\theta$8–$\theta$9 (Zhu et al., 2017).

A tomographic variant appears in reconstruction from the integrated Sachs–Wolfe effect. There the observable is the line integral

$p_o(\theta\mid\Lambda)$0

where $p_o(\theta\mid\Lambda)$1 solves a linearized wave equation for the scalar potential $p_o(\theta\mid\Lambda)$2. The data operator $p_o(\theta\mid\Lambda)$3 is treated as a PDE-constrained X-ray transform, and the inversion is stabilized with parametrices and filtered back-projection in the visible annular region $p_o(\theta\mid\Lambda)$4 (Chung et al., 2 Jul 2025). Theorem 1.1 gives Sobolev stability,

$p_o(\theta\mid\Lambda)$5

and the discrete study reports $p_o(\theta\mid\Lambda)$6–$p_o(\theta\mid\Lambda)$7 relative error for the full-data case, $p_o(\theta\mid\Lambda)$8 for partial data with early-stopped LSQR, and $p_o(\theta\mid\Lambda)$9 with $p(t)$0-FISTA or an edge-preserving prior in the $p(t)$1 detector setting (Chung et al., 2 Jul 2025). These cosmological examples collectively show that observable-space reconstruction can mean local estimation on the sky, inversion in redshift coordinates, or tomography of line-integral data, but in each case the observable domain itself determines the admissible reconstruction.

4. Partial visibility, completion, and geometric priors in 3D vision

In geometric vision, observable-space reconstruction is often motivated by missing surfaces, occlusions, and one-sided sensing. “4DComplete” addresses non-rigid motion estimation beyond the visible surface by taking as input a partial shape and motion observation, extracting a 4D time-space embedding, and jointly inferring missing geometry and motion field with a sparse fully-convolutional network. Training uses the synthetic DeformingThings4D dataset, comprising 1972 animation sequences spanning 31 different animals or humanoid categories with dense 4D annotation. Reported findings are that the method reconstructs high-resolution volumetric shape and motion field from partial observation, learns an entangled 4D feature representation useful for both shape and motion estimation, yields more accurate and natural deformation than As-Rigid-As-Possible deformation, and generalizes to unseen real-world sequences (Li et al., 2021).

A more explicit construction of observable space appears in POP3D. For a dense voxel grid $p(t)$2, a silhouette mask $p(t)$3, a monocular depth estimate $p(t)$4, and camera geometry $p(t)$5, the per-view feasible set is

$p(t)$6

and the accumulated observable space after $p(t)$7 views is

$p(t)$8

The method then projects the carved visual hull into a new viewpoint, computes an outpainting mask $p(t)$9, fills the unseen region with a pretrained latent diffusion model conditioned on a coarse render and a text prompt, and refits a signed-distance field with photometric, surface, normal, and eikonal losses (Ryu et al., 2023). The implementation uses ZoeDepth, OmniData, and TRACER for monocular cues; VolSDF for the implicit surface; and DreamBooth personalization to stabilize identity across outpainted views. An ablation on camera interval reports that intervals below $\zeta_n^{(O)}$0 accumulate boundary artifacts, intervals above $\zeta_n^{(O)}$1 induce generic hallucination, and $\zeta_n^{(O)}$2 is a robust default (Ryu et al., 2023).

Category-level object completion in indoor scenes uses a different observable-space strategy. Partial point clouds $\zeta_n^{(O)}$3 are subcategorized by a unidirectional Chamfer distance

$\zeta_n^{(O)}$4

and representative objects are chosen by ray-based uncertainty derived from rendered weight distributions $\zeta_n^{(O)}$5 and entropy $\zeta_n^{(O)}$6 (Lee et al., 2024). Objects are aligned to the representative with a unidirectional Chamfer alignment loss and then normalized to NOCS before training one CodeNeRF-style model per subcategory. On Replica room-2, the reported object-level results improve from $\zeta_n^{(O)}$7 cm Chamfer and $\zeta_n^{(O)}$8 completion ratio for vMAP* to $\zeta_n^{(O)}$9 cm and $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$0 for the category-level model, and ScanNet results are described as reducing completion error by $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$1–$dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$2 over instance-only baselines (Lee et al., 2024). This suggests that, in vision, observable-space reconstruction is often inseparable from completion priors: the observable region anchors the geometry, while category structure regularizes the unobserved part.

5. Observer-based and algebraic reconstructions from traces and temporal signals

A dynamical interpretation appears in phase-space reconstruction for lane intrusion action recognition. A tracked object is converted into a one-dimensional motion time series

$dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$3

where $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$4 is the object center and $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$5 are adjacent lane markings (Zhang et al., 2021). Instead of a hand-crafted delay embedding $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$6, PSRNet learns branches

$dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$7

for orders $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$8, supervises each branch with a mean-squared reconstruction loss $dN_a/d\theta(\Lambda)=N_a p_a(\theta\mid\Lambda)$9, concatenates the resulting latent descriptors, and classifies them with a 1D-CNN plus softmax using a combined loss $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$0 with $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$1 (Zhang et al., 2021). On the THU-IntrudBehavior dataset, using $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$2, $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$3, $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$4, Adam at $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$5, batch $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$6, and $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$7 epochs, PSRNet reaches $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$8 accuracy in 3-fold cross-validation, exceeding LSTM-, GRU-, TCN-, and MLP-based baselines (Zhang et al., 2021). Here the “observable” is not spatial but temporal: a scalar motion trace is reconstructed into a latent phase-space trajectory.

Observer-based geometric reconstruction with moving plenoptic cameras uses a continuous-time gradient descent on depth. The unknown state is the depth map $$p(\mathrm{det}\mid\theta)\equiv \int_{d>\mathrm{threshold}} p(d\mid\theta)\,dd,$$9, the measurement is the two-plane light field $s$00, and the update law is

$s$01

with Lyapunov candidate $s$02 for the depth error $s$03 (O'Brien et al., 2018). Under convexity and sign conditions on $s$04, $s$05, giving asymptotic convergence. In simulation with a textured sphere and a Lissajous-type 3D path over 5000 frames, the mean depth error drops from $s$06 m to $s$07 m in $s$08 frames with gain $s$09 (O'Brien et al., 2018). The explicit role of motion is observability: temporal parallax enlarges the set of lenslets that constrain each point.

At the operator level, observable-space reconstruction becomes algebraic tomography. For a non-Hermitian Floquet monodromy matrix $s$10 and an observable $s$11, the basic data are

$s$12

and Cayley–Hamilton implies the recurrence

$s$13

where $s$14 is the characteristic polynomial (Kamata, 23 May 2026). The observable resolvent

$s$15

separates a common spectral skeleton from observable-dependent dressing, and a Hankel null-vector recovers the coefficients $s$16. Multi-observable extensions and Liouville-space realizations generalize the construction to MIMO tomography. The crucial identifiability statement is that restricted observable algebras determine only a visible representative $s$17, while exact symmetries leave residual invisible sectors (Kamata, 23 May 2026). This is one of the clearest formal articulations of the boundary of observable-space reconstruction: the reconstruction is exact only within the visible operator subspace.

6. Identifiability, stability, and recurring limitations

A central theme across domains is that reconstruction quality is governed less by algorithmic expressivity alone than by the size and conditioning of the visible region. In the Floquet setting, observable restrictions produce a “visible representative” rather than the full operator; micromotion can enlarge the sampled operator space, but exact commuting symmetries enforce residual invisible sectors (Kamata, 23 May 2026). In ISW tomography, the visible domain is an annulus, and very sparse detector configurations recover the field only in the central annulus, not in outer or inner blind spots (Chung et al., 2 Jul 2025). In redshift-space reconstruction, only the gradient $s$18-mode part of $s$19 is reconstructed, while the vector $s$20-mode part and shell-crossing residuals appear as noise (Zhu et al., 2017).

A second recurring issue is that staying in observable space does not remove the need for calibration or nuisance modeling. In gravitational-wave inference, $s$21 enters denominators, so samples with very small detectability can cause instabilities; the stated remedies are to verify an accuracy threshold, remove unreliable samples, or increase emulator precision in the posterior support region (Toubiana et al., 17 Jul 2025). In CMB lensing, a finite-support filter introduces a multiplicative form factor $s$22 at large $s$23, requiring numerical deconvolution (Bucher et al., 2010). In POP3D, outpainting quality depends on viewpoint scheduling and personalization, with both too-small and too-large camera intervals degrading consistency (Ryu et al., 2023).

Vision systems built from real imagery add further observability bottlenecks. A pipeline for monocular reconstruction of non-cooperative space objects reports that segmentation-based background removal is essential for successful camera pose estimation: after masking, COLMAP registered $s$24 of extracted frames, whereas unmasked processing registered only a handful. The same pipeline introduces per-frame photometric correction,

$s$25

to model exposure variation, vignetting, color correction, and camera response, and finds that shadowed regions remain under-recovered while complex structures can be over-darkened or warm-tinted (Gopu et al., 30 Apr 2026). This illustrates a general point: observable-space reconstruction is often only as reliable as its preprocessing of masks, photometric consistency, or detector response.

A plausible implication is that the principal advantage of observable-space reconstruction is epistemic discipline. By conditioning on what is actually seen, detected, or sampled, these methods avoid unsupported extrapolation. The corresponding cost is equally consistent across the literature: reconstruction is confined to what the observable model can certify, and every domain retains a formally stated remainder—masked sky modes, shell-crossed structures, occluded geometry, sub-threshold source regions, or invisible operator sectors.